Changelog

Notable changes to VoiceLens, newest first: new analysis signals, integration features and documentation.

Notable product changes, newest first.

27 September 2026

Per-upload language: a single upload can now declare its own spoken language, overriding the channel's setting. Available as a picker on the dashboard upload page and as a top-level language field on every ingest variant.
Full transcription language list: the language pickers now offer every language the transcription engine supports (around 50, including Slovenian, Croatian, Serbian, Hebrew and Persian) instead of a curated nine. This matters because automatic detection covers only about 35 languages and labels the rest as the nearest language it knows — Slovenian audio was coming back as Bulgarian.
Fallback lexicon warning: the language badge on a call now warns when the call was transcribed in one language but graded with the English word-based detectors.

Public documentation launched at /docs, prerendered for search engines with locale-ready URLs, a Markdown twin of every page and an llms.txt index.
Solutions pages for five use cases and ten industries in six languages, with a Solutions menu in the public header and a footer that links to them.
Speaker attribution is now recorded per call (multichannel or diarized) and shown on the scorecard, so reviewers know whether "who said what" came from separate audio channels or was inferred.
Re-transcribe is available on calls whose transcript should be regenerated; it consumes minutes like a new upload.
Line-quality playback: clicking a line-quality event on the scorecard plays the affected part of the recording.
Re-evaluate now also measures line quality on calls analysed before the acoustic stage existed.
MP3 uploads are decoded in-process, so they no longer depend on ffmpeg being available.

Acoustic line quality: a new analysis pass measures signal-to-noise ratio, mains hum, hiss, clipping and dropouts, producing a 0 to 100 line-quality score and LINE_QUALITY info faults that never affect the agent's health score. Shown on the scorecard and exposed as lineQualityScore and snrDb.
Per-call language: calls carry their language, set on the channel or detected automatically, and fault lexicons are selected accordingly for English, German, Spanish, Italian, Arabic and Turkish.

Test calls: channels on Twilio can place scripted outbound test calls to an AI agent or IVR, record them and grade them like any other call. Test calls are flagged and excluded from analytics by default.

Agent identity on calls: agentId, agentName and agentType (ai or human) in the ingest metadata power per-agent analytics and bot-versus-human comparisons.
Rubric bootstrap: the first call in a workspace without a rubric automatically drafts one from the transcript for review.

Billing: Pro and Agency plans with Stripe Checkout and the customer portal; a usage page with minutes, AI drafts and a per-workspace breakdown.
Channel sampling: samplePercent and minDurationSec per channel to control which calls are graded and metered.
Workspace API keys (vlk_): read-only access to conversations, analytics and channels for your own reporting, with expiry, revocation and a 600 requests/minute limit.
Weekly digest: a Monday summary email per workspace with volume, health, top faults, busiest channels and alerts.
Provider credentials on channels: Twilio auth tokens are stored write-only and used to verify recording callbacks.
Alerts: rules on average health, danger-call rate, failure rate and call volume over a window of calls or hours, with email and webhook notifications and recovery events.

Signed webhooks with conversation.completed, conversation.failed and threshold.breached events, retries with backoff and delivery logs.
Embedded dashboards with per-embed tokens, themes, feature flags and allowed origins; the embed.js loader.
Shareable scorecards with optional transcript and audio, expiry and print support.
Golden set and calibration with mean absolute error, bias, tolerance share and correlation.
Rubric templates with slots, questions, compliance phrases, case branching, thresholds, detector toggles and versions; three built-in presets.
Sentiment trajectory and SENTIMENT_DROP detection for English calls.
Ingest API with multipart, from-URL and pre-signed uploads, idempotency and channel-scoped tokens.